Ninety Percent of AI in Manufacturing Is Not Software.

What thirty years in machine vision taught me about why most AI projects in manufacturing never make it to production.

It was almost midnight in a steel plant in Osaka, and I had been hunting the same fault for several days. The year was 1996. The system measured the ceramic lining inside vessels that carry molten steel, and if that lining wears too thin, the consequence is not a quality problem; it is a catastrophe. A laser scanned the inside of the vessel, thousands of single measurement points building a 3D point cloud of the lining surface, while a machine vision camera fixed the coordinate system between the scanner and the vessel by watching a plate mounted on the vessel's side, carrying four circular shapes attached with high precision. Those four circles told the system exactly where the vessel was, and every point in the cloud depended on that answer.

The answer was sometimes wrong, and the errors were random, with no pattern I could find. The fixture algorithm would locate the vessel perfectly a hundred times, then miss for no reason at all. The laser was fine. The code was correct. I checked it again, and it was still correct.

I have told the Osaka story before. I never told this part. Close to midnight, I finally saw it. Across the plant, a large crane was lifting the heavy vessels, and every lift sent a tremor through the structure. The whole building shook, not much; you could barely feel it standing on the floor. But the camera felt it, and when that small shake landed in exactly the moment the fixture algorithm was running, the coordinate system moved. The thousands of laser points were measured correctly. They were just placed in a frame of reference that had shifted. The measurements were right. The building was moving.

I went back to the hotel and wrote code through the night, two changes. First, a fixture algorithm built on the Best Linear Unbiased Estimator, a statistical method that combines noisy measurements into the most reliable estimate possible, weighting each piece of information by how much it can be trusted. Second, an indicator that sensed vibration during the fixture process, so the system knew when it could not trust its own eyes at all. I was back at the factory at six in the morning and updated the code. The customer wanted the acceptance test immediately, and it passed.

I have carried that night with me for thirty years, because it taught me where these systems actually fail. Not in the model. At the interface between the model and the physical world.

Today I keep reading about AI projects in manufacturing that never reach production. RAND studied this: more than 80 percent of AI projects fail, twice the failure rate of IT projects without AI, and almost every post-mortem reaches the same diagnosis: that the model was not good enough and needs to be replaced with a better one. That is almost never what actually happened. What happened is that the model was trained on data describing the process as documented, then it met the process as run. Those are two different factories. One exists in the quality manual. The other exists at two in the morning, when the humidity has shifted, the raw material comes from a different supplier, and a crane is lifting something heavy three bays over.

Anyone writing an AI budget this year should worry about one number. In my machine vision days, the algorithm was maybe ten percent of the work, and the other ninety percent was physics: lighting, calibration, fixturing, vibration. In the last three years, the ten percent became nearly free. Models that would have been a research project in 1996 are now an API call. So companies are buying the part that got cheap and ignoring the part that did not. The ninety percent is still the ninety percent. It has not gotten one cent cheaper, because it is not a software problem; it is an engineering problem, and it is solved by people who walk the floor and notice the crane.

I see the same law every day in diagnostics manufacturing. When you dispense reagents at picoliter scale, physics does not negotiate. Temperature, viscosity, and surface tension do not care what your model predicts. Any intelligence you put on top of a production line is only as good as the physical truth underneath it.

So my advice to anyone starting an AI project in manufacturing is simple. Do not start with the model, start with the camera mount. Audit your sensors before you audit your algorithms. Ask whether your data describes the process you run or the process you wrote down. Build your system so it knows when not to trust its own eyes. That vibration indicator from 1996 is still the most important design principle I know for intelligent machines. In 2026, the industry calls it uncertainty awareness. In Osaka, it was one night in a hotel and a statistics textbook I knew by heart.

The companies that get AI to work on the factory floor will not be the ones with the best models. They will be the ones who find the crane.

Kauko Väinämö, CEO, Ginolis.